I'm assuming you mean how you are supposed to label an unlabeled dataset. Why not use both? Use experts to (basically hand-label) a small dataset as bullying or non-bullying and then use a supervised learning model to train on this dataset to classify (sort) the a small unlabelled subset of the data into the 2 classes. After that you could manually check the classified data and clean up errors. Then use both datasets to train another model and keep the process going.

Generally speaking, there are two types of ML techniques: supervised and unsupervised.

In supervised techniques (classification), such as support vector machine, the model (algorithm) needs a pre-labelled dataset in order to learn from it, and therefore can predict the class labels of unseen samples based on what have been learned.

However, in unsupervised techniques (clustering), k-means clustering algorithm for example, the algorithm learns how to cluster samples without any predefined class labels, but needs some input parameters such as the number of clusters.

You must define classes (it is cyberbullying / it is not cyberbullying).

I'd first decide on the material, the object of classification, i.e. what would you like to classify: short texts like tweets and comments, or larger texts like posts in blogs, or maybe even images. Then, of course, you search for the items on the web that expose or, on the opposite, do not expose features of cyber-bullying. These items will be your labeled data set. After that you can apply different ML algorithms (not only SVM) to try and see which of them do a better classification of your data. In case you actually do that, don't forget to share your data publicly so that everyone could see that your research is valid ;)